Skip to content

Using PR Risk Scorer as an AI Coding Skill

You can integrate PR Risk Scorer into AI coding assistants so they score pull requests automatically as part of code reviews.

Claude Code

/score-pr slash command

The repository ships a custom Claude Code slash command at .claude/commands/score-pr.md. When you open the project in Claude Code, the /score-pr command becomes available.

Usage:

/score-pr

Claude will:

  1. Run git diff HEAD~1 HEAD to collect structural signals
  2. Score each dimension using the same weights as the GitHub Action
  3. Print a Markdown risk table with per-signal scores and a weighted total
  4. Recommend which signal to address first

Install for any project:

Copy .claude/commands/score-pr.md into the target repository's .claude/commands/ directory. The command is immediately available in Claude Code without configuration.

Global install:

To make the command available in every project, copy the file to your user-level commands directory:

mkdir -p ~/.claude/commands
cp .claude/commands/score-pr.md ~/.claude/commands/score-pr.md

GitHub Copilot (Codex / Chat)

Add the scorer's rubric to your workspace instructions so Copilot considers structural risk during code reviews.

Step 1 — Create workspace instructions

Create .github/copilot-instructions.md in your repository:

## PR review checklist

When reviewing a pull request, compute a structural risk score using these five dimensions:

| Signal | Weight | Method |
|--------|--------|--------|
| Diff Size | 20% | `min(total_lines / 1000 * 100, 100)` |
| Cyclomatic Complexity | 30% | Average radon CC score, normalized 1–26 → 0–100 |
| Coverage Gap | 20% | `(1 - line_rate) * 100` from coverage.xml |
| Dead Code | 15% | Vulture items, capped at 20 → 100 |
| Migrations | 15% | 0 files=0, 1=60, 2=80, 3+=100 |

Weighted total: `score = 0.20*diff + 0.30*complexity + 0.20*coverage + 0.15*dead + 0.15*migration`

Risk tiers: < 40 = LOW 🟢, 40–70 = MEDIUM 🟡, > 70 = HIGH 🔴

Include the score table in every PR review comment.

Step 2 — Use Copilot Chat

In VS Code, open Copilot Chat and ask:

@workspace Score this PR for structural risk using the rubric in copilot-instructions.md

Copilot will apply the weights and return a formatted risk table.

Cursor

Add the scoring rubric to your project's .cursor/rules file (or cursor_rules.md):

## PR Risk Scoring

Before approving any PR, compute and display a risk score:
- Diff Size (20%): lines changed / 1000, capped at 100
- Complexity (30%): average cyclomatic complexity via radon, normalized 0–100
- Coverage Gap (20%): 1 − coverage line-rate, as a percentage
- Dead Code (15%): vulture items / 20, capped at 100
- Migrations (15%): 0→0, 1→60, 2→80, 3+→100

Total = weighted sum. LOW < 40, MEDIUM 40–70, HIGH > 70.

OpenAI Codex / ChatGPT

You can attach the scoring rubric as a system prompt or a file upload. Create a plain-text file pr-risk-rubric.txt:

PR Risk Scoring Rubric

Signals and weights:
  Diff Size     20%  min(lines_changed / 1000 * 100, 100)
  Complexity    30%  min((avg_cc - 1) / 25 * 100, 100)   [radon output]
  Coverage Gap  20%  (1 - line_rate) * 100                [coverage.xml]
  Dead Code     15%  min(vulture_items / 20 * 100, 100)
  Migrations    15%  0→0, 1→60, 2→80, 3+→100

Total = weighted sum. Tiers: LOW < 40, MEDIUM 40–70, HIGH > 70.

For every PR, output a Markdown table with each signal's score and the weighted total.

Attach this file in the ChatGPT conversation or include it in your Codex system prompt.

GitHub Actions (automated)

The fastest way to get risk scores on every PR is the GitHub Action itself. See Getting Started for the workflow YAML.

The action posts the score as a PR comment and sets a commit status — no AI assistant required.

Next step: Configuration →